Protect Privacy and Client Data

Client trust is built partly on what you refuse to collect. A service can look efficient while creating unnecessary privacy risk through full mailbox access, unredacted transcripts, broad cloud folders, shared passwords, and AI prompts that receive more information than the task requires.

Privacy control is a design choice, not a paragraph added after delivery. For every stage of the workflow, ask what data is needed, why it is needed, who can see it, where it goes, how long it stays, and what happens if something goes wrong.

This chapter provides a practical control process. It is not a substitute for privacy law, contract review, or security advice for the specific engagement.

Build a Data Map

Start with an inventory that records five facts for each item:

• The data or file.
• Its purpose.
• Where it came from and what permission applies.
• Every approved place it goes.
• When it must be kept, returned, or deleted.

For the workshop example:

• Approved transcript: derive four customer moments; supplied by the buyer with confirmed authority; stored in the approved working folder and approved AI account; delete after final acceptance and the agreed record period.
• Program notes: verify dates and descriptions; supplied by the buyer; stored in the approved working folder; delete with the transcript.
• Attendee contact list: not needed for this scope; do not receive it; no approved destination; do not collect it.
• Reviewer comments: create the acceptance record; supplied by the named buyer reviewer; stored in the approved work record; retain only as the agreement requires.

The map should name actual files and fields, not broad categories such as "client documents." A broad description hides the risky details.

For each item, ask whether a smaller substitute would work: a redacted excerpt, a synthetic example, a count instead of individual records, or a manual review without uploading an entire file.

Use Data Minimization in Practice

Data minimization means collecting and using only what the task requires. Apply it before access:

• Request one transcript instead of the shared drive.
• Ask for the fields needed for the result.
• Redact identifiers that do not affect the work.
• Separate the person who needs access from people who merely want visibility.
• Replace raw examples with approved synthetic material when testing.

If the buyer sends extra material, do not assume the extra access is approved merely because it arrived. Pause, identify the unexpected content, and ask how it should be handled. Delete or return it when instructed through an appropriate process.

Establish the Permission Basis

The buyer must have the right to share the material, and you need an appropriate basis and agreement for receiving and processing it. Depending on the jurisdiction and situation, that may involve notice, consent, a business contract, confidentiality duties, or other requirements. The exact basis is a legal and organizational question.

Make the permission visible:

• The buyer confirms authority to supply the files.
• The engagement states the purpose and allowed processing.
• Data subjects receive any required notice.
• Subcontractors or software vendors are disclosed where required.
• AI processing is specifically approved where required.

Do not treat a general service contract as permission to use data for any future purpose. Separate delivery, internal improvement, marketing, and model training in your questions and controls.

Review AI Tools as Data Recipients

A tool receives data when you paste, upload, transcribe, or connect a file. Review:

• Who operates the service and where processing occurs.
• Which plan and settings apply to the account.
• Whether inputs or outputs may be used to train models.
• The default and configurable retention periods.
• Human access for support or review.
• Encryption, audit logs, and access controls available on the plan.
• Subprocessors, data locations, and incident terms.
• Export and deletion behavior.
• What the buyer contract and policy require.

Use current primary vendor documentation and the actual account settings, not a summary from an old article or an AI answer. Save the reviewed policy version or record when it was checked if the engagement warrants it.

Not every task belongs in an online AI tool. High-risk, regulated, highly identifiable, or contractually restricted data may require a local process, enterprise configuration, de-identification, or no AI at all.

Control Access by Role

Give each person the smallest access needed:

• The input coordinator can supply and withdraw source files.
• The service provider can process only the approved inputs.
• The reviewer can see drafts and source notes but perhaps not unrelated client files.
• The final approver can approve release.
• Other team members do not receive shared passwords or broad folder access.

Use named accounts where the system supports them. Avoid sending credentials through ordinary messages. Remove access when the engagement ends or when a person no longer needs it.

Access control also applies to local copies and backups. A secure cloud folder does not help if the same transcript sits in downloads, email attachments, sync folders, and old chat threads.

Design the Working Environment

Before work starts:

• Create a named engagement folder with access limited to the agreed people.
• Use approved devices and accounts.
• Turn on available security controls such as screen locks and multi-factor authentication.
• Keep source files, work files, and final files distinguishable.
• Avoid using personal messaging or unmanaged storage for client material.
• Use clear version names so the wrong file is not delivered.

If you work with contractors, define their data access and review duties before sending client files. A contract that says "confidential" but does not control access is only a statement of intention.

Set Retention and Deletion Rules

Define:

• When the working copy is no longer needed.
• Which acceptance or business records must remain.
• How long those records stay.
• Who approves exceptions.
• How deletion is performed across devices, tools, backups, and shared drives.
• How you confirm deletion when required.

The shortest period is not automatically lawful or operationally wise. The right period depends on the contract, dispute needs, tax rules, professional duties, privacy law, and the client's instructions. What matters is that the period is deliberate and followed.

Do not promise deletion you cannot execute. If a vendor retains data for a defined period, disclose or change the workflow.

Write the AI Handling Rule

Create a plain-language card for the engagement:

• Approved AI uses.
• Approved tools and accounts.
• Inputs allowed and prohibited.
• Required redaction before upload.
• Whether data may train models.
• Human review requirement.
• Vendor retention.
• Fallback if approval is withdrawn.
• Incident contact.

Example:

> Use the approved work account only to draft messages from the redacted transcript and approved notes. Do not upload attendee contact information, financial records, health information, credentials, or unrelated client files. Review every material claim against the source. If the tool displays an unexpected data request or the source contains prohibited data, stop and contact the coordinator.

An AI tool may produce a helpful draft and still create a privacy problem. The output-quality check and the data-handling check are separate.

Prepare for an Incident

Define what counts as a possible incident:

• Data sent to the wrong person.
• Unexpected personal data in a draft.
• Loss of a device or account.
• Broad or accidental folder access.
• Unapproved AI processing.
• A vendor notice or suspected unauthorized access.

Keep a short response card:

1. Stop the affected workflow.
2. Preserve essential facts without spreading the data further.
3. Notify the named internal contact.
4. Determine what data, people, systems, and time are involved.
5. Follow the contract and applicable legal notification or assessment process.
6. Correct root causes and record lessons.

Do not investigate by pasting the leaked material into another AI tool. Do not promise that an incident has no legal consequences before the responsible process reviews it.

A Worked Privacy Review

The following is an illustrative composite, not legal or security advice.

Nina receives a workshop transcript, a program notes file, and an unrequested spreadsheet containing attendee names, email addresses, and meal preferences. Her engagement allows one redacted transcript and one notes file. The buyer has confirmed authority to share them but has not asked for the spreadsheet.

Nina stops before uploading anything. She tells the coordinator that the spreadsheet is outside the approved input list and asks whether it was sent by mistake. The coordinator confirms the file is unnecessary and requests deletion. Nina deletes it from the download location and confirms the action through the agreed channel.

The transcript contains speaker labels and a few attendee names. The coordinator supplies a redacted copy and confirms which labels are needed to preserve meaning. Nina checks the approved copy is the one in her working folder.

The workflow card allows a designated AI account for drafting from redacted text. It prohibits upload of attendee details. The tool's plan has a defined retention period. Nina records the reviewed policy date and uses only the approved account.

During drafting, a generated message adds a surname that was not visible in the approved transcript. Nina treats that as a possible data or hallucination event. She does not publish the message. She checks the source, confirms there is no surname, removes it, and adds a name check to the workflow. Because no data appears to have left the approved environment, she records the issue as a quality exception and follows the agreed incident questions with the coordinator.

The pilot ends with a shorter transcript, a clearer working folder, and a documented deletion date for the source files.

Source Claims About Privacy

Privacy requirements change. Use current primary sources:

• The relevant privacy or consumer regulator.
• The buyer's contract and internal policy.
• The vendor's current data processing and retention terms.
• The account's actual administrative settings.
• Official guidance for your jurisdiction and service type.

Do not rely on "the tool says it is private" or an AI summary of the law. Distinguish the tool's security claims from your duties, the buyer's duties, and the terms that govern the specific data.

Your Chapter Exercise

For your proposed service:

1. Build the five-column data map.
2. Mark each item as necessary, optional, or prohibited.
3. Identify one safe substitute for unnecessary data.
4. List the permission questions the buyer must answer.
5. Review one AI tool's current data and retention terms.
6. Create the working-folder and access plan.
7. Write the AI handling card.
8. Define retention, return, and deletion steps.
9. Write a possible-incident response card.
10. Test the workflow with synthetic data and one intentionally prohibited field.

Chapter Deliverable

Create a privacy control sheet containing:

• Engagement purpose and permitted use.
• Data inventory and minimization decisions.
• Buyer and provider responsibilities.
• Tool and vendor register.
• Access roles and working-folder map.
• AI approval, redaction, and training-use rules.
• Retention, return, and deletion plan.
• Incident contact and response card.
• Primary sources reviewed and date checked.
• Stop conditions for unexpected or prohibited data.

Source and Decision Caution

Requirements for privacy, confidentiality, consent, security, breach response, cross-border transfer, records, and automated processing vary by jurisdiction and by the type of data. This chapter is an operating template, not legal, security, or compliance advice. Confirm the actual duties with current official guidance and qualified reviewers.

AI providers can change features, retention, accounts, and terms. Verify the current contract and settings before processing client data. Keep approvals and records in controlled locations, and do not use AI to summarize or generate the response to a suspected privacy incident without checking the handling requirements first.

The next chapter turns the approved scope, human checks, and privacy controls into the operating rhythm of one paid engagement.